Learning a Go Heuristic with Tilde

Abstract

In Go, an important factor that hinders search is the large branching factor, even in local problems. Human players are strong at recognizing frequently occurring shapes and vital points. This allows them to select the most promising moves and to prune the search tree. In this paper we argue that many of these shapes can be represented as relational concepts. We present an application of the relational learner TILDE in which we learn a heuristic that gives values to candidate-moves in tsume-go (life and death) problems. Such a heuristic can be used to limit the number of evaluated moves. Even if all moves are evaluated, alpha-beta search can be sped up considerably when the candidate-moves are approximately ordered from good to bad. We validate our approach with experiments and analysis.

Compact Representation of Knowledge Bases in Inductive Logic Programming

“…A more efficient alternative is to store only the initial state together with the sequence of actions. A similar situation is encountered when learning to play games like chess or Go, where the knowledge base consists of a number of played games (see Ramon, Francis, and Blockeel, 2000, for an example).…”

“…Go is an abstract two-person complete-information deterministic board game, popular in Asia. The Go data set (Ramon, Francis, and Blockeel, 2000) that we use here contains a log generated by an alpha-beta search algorithm. Starting from a number of initial states, the search algorithm explores the game state-tree and evaluates a number of moves for each state.…”

Enhancing understanding and improving prediction of severe weather through spatiotemporal relational learning

“…We give a brief overview here and the full details of the learning algorithm are provided in Appendix A . SRPTs differ from existing tree-based relational learning approaches such as TILDE (Blockeel and De Raedt 1998 ; Ramon et al 2002 ) in their ability to handle the discovery of multi-dimensional relationships (such as the spatial ones introduced here) and their ability to handle temporally varying data. SRRF also differs from the Relational Probability Tree (RPT Neville et al 2003 ) and the temporal extensions to the RPT (Sharan and Neville 2007 , 2008 ) in its ability to handle spatially and spatiotemporally varying relational data.…”

Single-Graph Support Measures

“…Over the last several years, graph mining has emerged as a new field within contemporary data mining. One of the central tasks is the search for subgraphs, called patterns, that occur frequently in either a collection of graphs (e.g., databases of molecules [6], game positions [15], scene descriptions) or in a single large graph Dehmer/Quantitative Graph Theory K19041_C010 Pageproof Page 303 2014-6-7 (e.g., the Internet, citation networks [16], social networks [12], protein interaction networks [10]). In the literature, the terms frequency and support have been used interchangeably to denote the measure to quantify the prevalence of a pattern.…”